Two-stage deep-learning-based colonoscopy polyp detection incorporating fisheye and reflection correction

Colonoscopy is a useful method for the diagnosis and management of colorectal diseases. Many computer-aided systems have been developed to assist clinicians in detecting colorectal lesions by analyzing colonoscopy images. However, fisheye-lens distortion and light reflection in colonoscopy images can substantially affect the clarity of these images and their utility in detecting polyps. This study proposed a two-stage deep-learning model to correct distortion and reflections in colonoscopy images and thus facilitate polyp detection.

Images were collected from the PolypSet dataset, the Kvasir-SEG dataset, and one medical center's patient archiving and communication system. The training, validation, and testing datasets comprised 808, 202, and 1100 images, respectively. The first stage involved the correction of fisheye-related distortion in colonoscopy images and polyp detection, which was performed using a convolutional neural network. The second stage involved the use of generative and adversarial networks for correcting reflective colonoscopy images before the convolutional neural network was used for polyp detection.

The model had higher accuracy when it was validated using corrected images than when it was validated using uncorrected images (96.8% vs 90.8%, P < 0.001). The model's accuracy in detecting polyps in the Kvasir-SEG dataset reached 96%, and the area under the receiver operating characteristic curve was 0.94.

The proposed model can facilitate the clinical diagnosis of colorectal polyps and improve the quality of colonoscopy.